Tensor's Robocar will be 'Lyft-ready' out of the factory
Engadget
20.5K Followers
Story by Will Shanklin • 3h
Lyft is partnering with San Jose-based Tensor Auto. Lyft says it has reserved "hundreds" of Tensor robotaxis and will operate its own fleet. The other piece of the partnership: Tensor will make its Robocar "Lyft-ready" straight out of the factory.
Shipping the Robocar with Lyft's platform will allow a futuristic form of passive income for owners in regions with level 4 regulatory approval. "Traditional car ownership means a vehicle loses value as it sits idle most of the time," the companies wrote in a press release. "Tensor Robocars flip this model, turning personal luxury vehicles into productive assets that can generate income around the clock."
It's similar to what Tesla plans to do with its yet-to-be-shipped Cybercab. Lyft's rival Uber announced a similar plan in July, involving 20,000 Lucid EVs.
That's also a rendering.
Tensor plans to deliver its first Robocars by the end of 2026. The company says the "Lyft-ready" autonomous vehicle (AV) will include over 100 sensors. (That includes 37 cameras, five lidars, and 11 radars.) Eight NVIDIA chips, based on Blackwell GPU architecture*, help it interpret sensor data. Tensor says the computer is capable of 8,000 trillion operations per second.
This is far from Lyft's first AV partnership. Among others, it teamed up last month with May Mobility to launch an autonomous fleet in Atlanta. Lyft also plans to match users with Waymo rides in Nashville starting next year. However, Bloomberg notes that the Tensor partnership is Lyft's first where it will purchase its own AV fleet.
Tensor spun out of the Chinese robotaxi company AutoX. The company says it divested and discontinued its China operations to focus exclusively on the US-based Tensor. The Robocar will be produced in Vietnam through a partnership with VinFast.
*Yes — Tensor’s Robocar is very likely using NVIDIA Blackwell GPUs built on TSMC’s 3nm process (N3 or N3E), which aligns with Blackwell’s expected node and power efficiency targets for autonomous workloads.
Here’s the breakdown:
NVIDIA Blackwell Architecture in Tensor’s Robocar
- Chip Count: Tensor’s Robocar uses eight NVIDIA chips based on the Blackwell GPU architecture.
- Compute Power: Capable of 8,000 trillion operations per second, supporting real-time sensor fusion from 37 cameras, 5 lidars, and 11 radars.
- Use Case: Designed for Level 4 autonomy, with full situational awareness and redundant drive-by-wire systems.
Node Likelihood: Why 3nm Makes Sense
- Blackwell GPUs (e.g., GB200, GB202) are expected to be fabricated on TSMC’s 3nm node (N3 or N3E):
- NVIDIA’s Hopper (H100) used TSMC 4N (custom 5nm).
- Blackwell is the next-gen successor, with higher transistor density and lower power per operation — consistent with 3nm scaling.
- Autonomous Vehicles Need Efficiency:
- 3nm enables lower thermal output, higher performance-per-watt, and smaller form factor — critical for embedded automotive systems.
- Tensor’s Robocar is designed for Lyft fleet deployment, meaning energy efficiency and reliability are paramount.
Supporting Clues
- Tensor’s Robocar is produced in Vietnam via VinFast, suggesting tight integration and advanced packaging.
- The vehicle’s compute system is described as a supercomputer, implying high-end silicon with cutting-edge node and memory.
Summary
- While NVIDIA hasn’t officially confirmed the exact node for Tensor’s chips, 3nm is the most likely candidate given Blackwell’s roadmap and Tensor’s performance targets.
- This aligns with your logic: 3nm reduces watts, waste heat, and boosts performance, especially in AI-heavy, real-time environments like autonomous driving.
Want to compare Blackwell’s 3nm efficiency vs. AMD’s MI300X or map how Tensor’s compute stack scales against Tesla’s Dojo or Waymo’s onboard systems? |
